condition variable per context instead of one mutex and one conditional
variable for the whole subsystem.
This should increase concurrency in this subsystem.
It also opens the way for further changes which are necessary to solve
such bugs as bug #46272 "MySQL 5.4.4, new MDL: unnecessary deadlock"
and bug #37346 "innodb does not detect deadlock between update and alter
table".
Two other notable changes done by this patch:
- MDL subsystem no longer implicitly acquires global intention exclusive
metadata lock when per-object metadata lock is acquired. Now this has
to be done by explicit calls outside of MDL subsystem.
- Instead of using separate MDL_context for opening system tables/tables
for purposes of I_S we now create MDL savepoint in the main context
before opening tables and rollback to this savepoint after closing
them. This means that it is now possible to get ER_LOCK_DEADLOCK error
even not inside a transaction. This might happen in unlikely case when
one runs DDL on one of system tables while also running DDL on some
other tables. Cases when this ER_LOCK_DEADLOCK error is not justified
will be addressed by advanced deadlock detector for MDL subsystem which
we plan to implement.
This change is supposed to reduce number of ER_LOCK_DEADLOCK
errors which occur when multi-statement transaction encounters
conflicting metadata lock in cases when waiting is possible.
The idea is not to fail ER_LOCK_DEADLOCK error immediately when
we encounter conflicting metadata lock. Instead we release all
metadata locks acquired by current statement and start to wait
until conflicting lock go away. To avoid deadlocks we use simple
empiric which aborts waiting with ER_LOCK_DEADLOCK error if it
turns out that somebody is waiting for metadata locks owned by
this transaction.
This patch also fixes bug #46273 "MySQL 5.4.4 new MDL: Bug#989
is not fully fixed in case of ALTER".
The bug was that concurrent execution of UPDATE or MULTI-UPDATE
statement as a part of multi-statement transaction that already
has used table being updated and ALTER TABLE statement might have
resulted of loss of isolation between this transaction and ALTER
TABLE statement, which manifested itself as changes performed by
ALTER TABLE becoming visible in transaction and wrong binary log
order as a consequence.
This problem occurred when UPDATE or MULTI-UPDATE's wait in
mysql_lock_tables() call was aborted due to metadata lock
upgrade performed by concurrent ALTER TABLE. After such abort all
metadata locks held by transaction were released but transaction
silently continued to be executed as if nothing has happened.
We solve this problem by changing our code not to release all
locks in such case. Instead we release only locks which were
acquired by current statement and then try to reacquire them
by restarting open/lock tables process. We piggyback on simple
deadlock detector implementation since this change has to be
done anyway for it.
Postfix for Bug#48210 FLUSH TABLES WITH READ LOCK deadlocks
against concurrent CREATE PROCEDURE
Rewrote the second test to use DROP PROCEDURE instead of
CREATE USER as CREATE USER does not work with embedded server.
Bug #48210 FLUSH TABLES WITH READ LOCK deadlocks
against concurrent CREATE PROCEDURE
This deadlock occured between
a) CREATE PROCEDURE (or other commands listed below)
b) FLUSH TABLES WITH READ LOCK
If the execution of them happened in the following order:
- a) opens a table (e.g. mysql.proc)
- b) locks the global read lock (or GRL)
- a) sleeps inside wait_if_global_read_lock()
- b) increases refresh_version and sleeps waiting
for old tables to go away
Note that a) must start waiting on the GRL before FLUSH increases
refresh_version. Otherwise a) won't wait on the GRL and instead
close its tables for reopen, allowing FLUSH to complete and thus
avoid the deadlock.
With this patch the deadlock is avoided by making CREATE PROCEDURE
acquire a protection against global read locks before it starts
executing. This means that FLUSH TABLES WITH READ LOCK will have
to wait until CREATE PROCEDURE completes before acquiring the global
read lock, thereby avoiding the deadlock.
This is implemented by introducing a new SQL command flag called
CF_PROTECT_AGAINST_GRL. Commands marked with this flag will
acquire a GRL protection in the beginning of mysql_execute_command().
This patch adds the flag to CREATE, ALTER and DROP for PROCEDURE
and FUNCTION, as well as CREATE USER, DROP USER, RENAME USER and
REVOKE ALL. All these commands either call open_grant_tables() or
open_system_table_for_updated() which make them susceptible for
this deadlock.
The patch also adds the CF_PROTECT_AGAINST_GRL flag to a number
of commands that previously acquired GRL protection in their
respective SQLCOM case in mysql_execute_command().
Test case that checks for GRL protection for CREATE PROCEDURE
and CREATE USER added to mdl_sync.test.
Bug#42546 Backup: RESTORE fails, thinking it finds an existing table
The problem occured when a MDL locking conflict happened for a non-existent
table between a CREATE and a INSERT statement. The code for CREATE
interpreted this lock conflict to mean that the table existed,
which meant that the statement failed when it should not have.
The problem could occur for CREATE TABLE, CREATE TABLE LIKE and
ALTER TABLE RENAME.
This patch fixes the problem for CREATE TABLE and CREATE TABLE LIKE.
It is based on code backported from the mysql-6.1-fk tree written
by Dmitry Lenev. CREATE now uses normal open_and_lock_tables() code
to acquire exclusive locks. This means that for the test case in the bug
description, CREATE will wait until INSERT completes so that it can
get the exclusive lock. This resolves the reported bug.
The patch also prohibits CREATE TABLE and CREATE TABLE LIKE under
LOCK TABLES. Note that this is an incompatible change and must
be reflected in the documentation. Affected test cases have been
updated.
mdl_sync.test contains tests for CREATE TABLE and CREATE TABLE LIKE.
Fixing the issue for ALTER TABLE RENAME is beyond the scope of this
patch. ALTER TABLE cannot be prohibited from working under LOCK TABLES
as this could seriously impact customers and a proper fix would require
a significant rewrite.
------------------------------------------------------------
revno: 2617.68.10
committer: Dmitry Lenev <dlenev@mysql.com>
branch nick: mysql-next-bg46673
timestamp: Tue 2009-09-01 19:57:05 +0400
message:
Fix for bug #46673 "Deadlock between FLUSH TABLES WITH READ LOCK and DML".
Deadlocks occured when one concurrently executed transactions with
several statements modifying data and FLUSH TABLES WITH READ LOCK
statement or SET READ_ONLY=1 statement.
These deadlocks were introduced by the patch for WL 4284: "Transactional
DDL locking"/Bug 989: "If DROP TABLE while there's an active transaction,
wrong binlog order" which has changed FLUSH TABLES WITH READ LOCK/SET
READ_ONLY=1 to wait for pending transactions.
What happened was that FLUSH TABLES WITH READ LOCK blocked all further
statements changing tables by setting global_read_lock global variable
and has started waiting for all pending transactions to complete.
Then one of those transactions tried to executed DML, detected that
global_read_lock non-zero and tried to wait until global read lock will
be released (i.e. global_read_lock becomes 0), indeed, this led to a
deadlock.
Proper solution for this problem should probably involve full integration
of global read lock with metadata locking subsystem (which will allow to
implement waiting for pending transactions without blocking DML in them).
But since it requires significant changes another, short-term solution
for the problem is implemented in this patch.
Basically, this patch restores behavior of FLUSH TABLES WITH READ LOCK/
SET READ_ONLY=1 before the patch for WL 4284/bug 989. By ensuring that
extra references to TABLE_SHARE are not stored for active metadata locks
it changes these statements not to wait for pending transactions.
As result deadlock is eliminated.
Note that this does not change the fact that active FLUSH TABLES WITH
READ LOCK lock or SET READ_ONLY=1 prevent modifications to tables as
they also block transaction commits.
------------------------------------------------------------
revno: 2617.68.7
committer: Dmitry Lenev <dlenev@mysql.com>
branch nick: mysql-next-bg46044
timestamp: Thu 2009-08-27 10:22:17 +0400
message:
Fix for bug #46044 "MDL deadlock on LOCK TABLE + CREATE TABLE HIGH_PRIORITY
FOR UPDATE".
Deadlock occured when during execution of query to I_S we tried to open
a table or its .FRM in order to get information about it and had to wait
because we have encountered exclusive metadata lock on this table held by
a DDL operation from another connection which in its turn waited for some
resource currently owned by connection executing this I_S query.
For example, this might have happened if one under LOCK TABLES executed I_S
query targeted to particular table (which was not among locked) and also
concurrently tried to create this table using CREATE TABLE SELECT which
had to wait for one of tables locked by the first connection.
Another situation in which deadlock might have occured is when I_S query,
which was executed as part of transaction, tried to get information about
table which just has been dropped by concurrent DROP TABLES executed under
LOCK TABLES and this DROP TABLES for its completion also had to wait
transaction from the first connection.
This problem stemmed from the fact that opening of tables/.FRMs for I_S
filling is happening outside of connection's main MDL_context so code
which tries to detect deadlocks due to conflicting metadata locks doesn't
work in this case. Indeed, this led to deadlocks when during I_S filling
we tried to wait for conflicting metadata lock to go away, while its owner
was waiting for some resource held by connection executing I_S query.
This patch solves this problem by avoiding waiting in such situation.
Instead we skip this table and produce warning that information about
it was omitted from I_S due to concurrent DDL operation. We still wait
for conflicting metadata lock to go away when it is known that deadlock
is not possible (i.e. when connection executing I_S query does not hold
any metadata or table-level locks).
Basically, we apply our standard deadlock avoidance technique for metadata
locks to the process of filling of I_S tables but replace ER_LOCK_DEADLOCK
error with a warning.
Note that this change is supposed to be safe for 'mysqldump' since the
only its mode which is affected by this change is --single-transaction mode
is not safe in the presence of concurrent DDL anyway (and this fact is
documented). Other modes are unaffected because they either use
SHOW TABLES/SELECT * FROM I_S.TABLE_NAMES which do not take any metadata
locks in the process of I_S table filling and thus cannot skip tables or
execute I_S queries for tables which were previously locked by LOCK TABLES
(or in the presence of global read lock) which excludes possibility of
encountering conflicting metadata lock.
------------------------------------------------------------
revno: 2617.69.37
committer: Dmitry Lenev <dlenev@mysql.com>
branch nick: mysql-next-bg46748
timestamp: Fri 2009-08-21 18:17:02 +0400
message:
Fix for bug #46748 "Assertion in MDL_context::wait_for_locks()
on INSERT + CREATE TRIGGER".
Concurrent execution of statements involving stored functions or triggers
which were using several tables and DDL statements which affected those
tables on debug build of server might have led to assertion failures in
MDL_context::wait_for_locks(). Non-debug build was not affected.
The problem was that during back-off which happens when open_tables()
encounters conflicting metadata lock for one of the tables being open
we didn't reset MDL_request::ticket value for requests which correspond
to tables from extended prelocking set. Since these requests are part
of of list of requests to be waited for in Open_table_context this broke
assumption that ticket value for them is 0 in MDL_context::wait_for_locks()
and caused assertion failure.
This fix ensures that close_tables_for_reopen(), which performs this back-off
resets MDL_request::ticket value not only for tables directly used by the
statement but also for tables from extended prelocking set, thus satisfying
assumption described above.
------------------------------------------------------------
revno: 2617.23.18
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Mon 2009-03-02 18:18:26 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
This is a prerequisite patch:
These changes are intended to split lock requests from granted
locks and to allow the memory and lifetime of granted locks to
be managed within the MDL subsystem. Furthermore, tickets can
now be shared and therefore are used to satisfy multiple lock
requests, but only shared locks can be recursive.
The problem is that the MDL subsystem morphs lock requests into
granted locks locks but does not manage the memory and lifetime
of lock requests, and hence, does not manage the memory of
granted locks either. This can be problematic because it puts the
burden of tracking references on the users of the subsystem and
it can't be easily done in transactional contexts where the locks
have to be kept around for the duration of a transaction.
Another issue is that recursive locks (when the context trying to
acquire a lock already holds a lock on the same object) requires
that each time the lock is granted, a unique lock request/granted
lock structure structure must be kept around until the lock is
released. This can lead to memory leaks in transactional contexts
as locks taken during the transaction should only be released at
the end of the transaction. This also leads to unnecessary wake
ups (broadcasts) in the MDL subsystem if the context still holds
a equivalent of the lock being released.
These issues are exacerbated due to the fact that WL#4284 low-level
design says that the implementation should "2) Store metadata locks
in transaction memory root, rather than statement memory root" but
this is not possible because a memory root, as implemented in mysys,
requires all objects allocated from it to be freed all at once.
This patch combines review input and significant code contributions
from Konstantin Osipov (kostja) and Dmitri Lenev (dlenev).
